Message 1: Start Task A
           ↓ wait for completion
Message 2: Start Task B
           ↓ wait for completion
Message 3: Start Task C
           ↓ wait for completion

Total time = A + B + C = 90 seconds (if each takes 30s)

Message 1: Start Task A ─┐
           Start Task B ─┼─ All run simultaneously
           Start Task C ─┘

Total time ≈ max(A, B, C) = 30 seconds

<!-- CORRECT: All tasks in single function_calls block = PARALLEL -->
<function_calls>
<invoke name="runSubagent">
<parameter name="description">Analyze authentication module</parameter>
<parameter name="prompt">Review src/auth for security patterns...</parameter>
</invoke>
<invoke name="runSubagent">
<parameter name="description">Analyze API layer</parameter>
<parameter name="prompt">Review src/api for REST best practices...</parameter>
</invoke>
<invoke name="runSubagent">
<parameter name="description">Analyze database layer</parameter>
<parameter name="prompt">Review src/db for query optimization...</parameter>
</invoke>
</function_calls>

Implementation Plan:
1. Implement auth module
2. Create API endpoints
3. Add database schema
4. Write unit tests
5. Update documentation

Launch 5 parallel subagents:
├─ Subagent 1: Implement auth module
├─ Subagent 2: Create API endpoints
├─ Subagent 3: Add database schema
├─ Subagent 4: Write unit tests
└─ Subagent 5: Update documentation

Codebase Structure:
├── src/auth/
├── src/api/
├── src/db/
└── src/ui/

Launch 4 parallel subagents:
├─ Subagent 1: Analyze src/auth
├─ Subagent 2: Analyze src/api
├─ Subagent 3: Analyze src/db
└─ Subagent 4: Analyze src/ui

Code Review Perspectives:
- Security vulnerabilities
- Performance bottlenecks
- Test coverage gaps
- Architecture patterns

Launch 4 parallel subagents:
├─ Subagent 1: Security review
├─ Subagent 2: Performance analysis
├─ Subagent 3: Test coverage review
└─ Subagent 4: Architecture assessment

Verification Subagents (all parallel):
├─ Syntax & Type Checker
├─ Test Runner
├─ Lint & Style Checker
├─ Security Scanner
└─ Build Validator

Then (sequential, after above complete):
├─ False Positive Filter
├─ Missing Issues Finder
└─ Context Validator

{
  "todos": [
    { "content": "Analyze auth module", "status": "in_progress", "activeForm": "Analyzing auth module" },
    { "content": "Analyze API layer", "status": "in_progress", "activeForm": "Analyzing API layer" },
    { "content": "Analyze database layer", "status": "in_progress", "activeForm": "Analyzing database layer" },
    { "content": "Synthesize findings", "status": "pending", "activeForm": "Synthesizing findings" }
  ]
}

{
  "todos": [
    { "content": "Analyze auth module", "status": "completed", "activeForm": "Analyzing auth module" },
    { "content": "Analyze API layer", "status": "completed", "activeForm": "Analyzing API layer" },
    { "content": "Analyze database layer", "status": "in_progress", "activeForm": "Analyzing database layer" },
    { "content": "Synthesize findings", "status": "pending", "activeForm": "Synthesizing findings" }
  ]
}

WRONG (Sequential):
Message 1: "I'll start analyzing the auth module..."
           <invoke name="runSubagent">Analyze auth</invoke>
Message 2: "Now let me analyze the API..."
           <invoke name="runSubagent">Analyze API</invoke>

RIGHT (Parallel):
Message 1: "I'll analyze all modules in parallel..."
           <function_calls>
           <invoke name="runSubagent">Analyze auth</invoke>
           <invoke name="runSubagent">Analyze API</invoke>
           <invoke name="runSubagent">Analyze DB</invoke>
           </function_calls>

WRONG:
"I'm going to launch three parallel tasks to analyze the codebase."
[waits for response]
"Now launching the tasks..."

RIGHT:
"Launching three parallel analysis tasks now:"
<function_calls>
<invoke name="runSubagent">...</invoke>
<invoke name="runSubagent">...</invoke>
<invoke name="runSubagent">...</invoke>
</function_calls>

WRONG:
Just dump all task outputs without integration

RIGHT:
After receiving all results, synthesize:
- Identify common themes
- Resolve contradictions
- Prioritize findings
- Create unified recommendations

undefined

---

RULE #1:
  ALL runSubagent calls in SINGLE function_calls block = PARALLEL
  runSubagent calls in SEPARATE messages = SEQUENTIAL

PATTERNS:
  Task-based:       One subagent per task
  Directory-based:  One subagent per directory
  Perspective-based: One subagent per viewpoint
  Adversarial:      Multiple competing reviewers

TODOLIST:
  Mark ALL parallel tasks as in_progress BEFORE launching
  Mark each as completed AFTER receiving results

SPEEDUP:
  N parallel tasks ≈ Nx faster
  (5 tasks @ 30s each: 150s → 30s)

CHECKLIST:
  ☐ Tasks independent?
  ☐ No shared files?
  ☐ No dependencies?
  ☐ All in ONE function_calls block?
  ☐ Synthesis planned?